Table 3 HVAE training configuration. All parameter settings that were used for training the HVAE.

Latent dimensions (d)

2

Negative curvature (c)

1.0

Reconstruction loss weight (\(W_{\text {recon}}\))

30.0

KL-divergence loss weight (\(W_{\text {KL}}\))

1.0

Regularisation loss weight (\(W_{\text {reg}}\))

50.0

Hyper-KL-divergence loss weight (\(W_{\text {KL,hyp}}\))

1000.0

Hyper-regularisation loss weight (\(W_{\text {reg,hyp}}\))

10000.0

Group-level variance (\(\varvec{\sigma }_g\))

0.1

Base dispersion (s)

2.0

Encoder residual blocks (\(N_{\text {encoder}}\))

1

Encoder weight proportion vector (\(\textbf{P}_{\text {encoder}}\))

[1, 0.95, 0.9, 0.85]

Decoder residual blocks (\(N_{\text {decoder}}\))

2

Decoder weight proportion vector (\(\textbf{P}_{\text {decoder}}\))

[1.0, 0.95, 0.9, 0.8, 0.7, 0.6, 0.5]

Minimum epochs

300

Maximum epochs

500